Experienced Weight Stigma Changes Following Bariatric Surgery: a Systematic Review and Meta-Analysis.

It is of continuing interest to identify demographic and clinical features that are associated with late onset of dementia, especially if they are causal and modifiable. Such features may also help identify the best candidates for clinical trials of new therapies, e.g., those who are as yet unaffected but at high risk. This analysis is best done in large cohorts so that interactions and other potentially rare features may be identified. One such cohort is the National Alzheimer’s Coordinating Center (NACC) database, which contains clinical, neuropsychological and demographic data from thousands of participants enrolled in Alzheimer’s Disease Research Centers across the United States (). Because screening for dementia is enhanced for NACC study participants, we focus our attention on those who experienced onset of dementia prior to their entry to NACC, during which time they underwent community-based screening. These analyses must account for the fact that subjects were alive and under surveillance at entry to NACC. This translates into a sequence of events; cognitive decline precedes NACC entry, which precedes death or end of follow-up. We previously termed this framework (). Estimators have been derived for the distribution function of simple right-truncated data (e.g., ; ), as well as for regression models (e.g., ). For sequentially-truncated data, proposed nonparametric maximum likelihood estimators (NPMLE’s) under two models of dependence. considered a related problem of two correlated event times, one of which is subject to right truncation, and derived a nonparametric bivariate estimator. Regression modeling in related settings of sequential truncation has been addressed in the literature on multi-state models, typically assuming Markov or semi-Markov transitions, often without right truncation (; ). Within the framework of , proposed a semiparametric linear transformation regression model (e.g., Cox) for bivariate correlated event times under strong assumptions of unconditional independence. In this paper, we derive computationally tractable methods for regression in the setting of sequential truncation that require no, or weaker, assumptions on independence among event times and covariates, allow for right truncation, and apply to both Cox and accelerated failure time models. We do this using the simple jackknife pseudo-observation regression approach () building on our NPMLE for this setting (). A related strategy was taken by in a multi-state setting without right truncation under a restrictive unconditional independence assumption. Pseudo-observations automatically account for the complex truncation through calculation of each individual’s contribution to the estimated marginal survival function. We show that this approach is biased in some settings, as it is for left truncation (). We thus propose a modified pseudo-observation approach, which reduces to a weighted binomial regression. We have expanded our R package, (), to include regression modeling under sequential truncation.